Frame-Based Slip Detection for an Underactuated Robotic Gripper for Assistance of Users with Disabilities

Stable grasping is essential for assistive robots aiding individuals with severe motor–
sensory disabilities in their everyday lives. Slip detection can prevent unstably grasped objects from
falling out of the gripper and causing accidents. Recent research on slip detection focuses on tactile
sensing; however, not every robot arm can be equipped with such sensors. In this paper, we propose
a slip detection method solely based on data collected by a RealSense D435 Red Green Blue-Depth
(RGBd) camera. By utilizing Farneback optical flow (OF) to estimate the motion field of the grasped
object relative to the gripper, while also removing potential background noise, the algorithm can
perform in a multitude of environments. The algorithm was evaluated on a dataset of 28 daily objects
that were lifted 30 times each, resulting in a total of 840 frame sequences. Our proposed slip detection
method achieves an accuracy of up to 82.38% and a recall of up to 87.14%, which is comparable
to state-of-the-art approaches when only using camera data. When excluding objects for which
movements are challenging for vision-based methods to detect, such as untextured or transparent
objects, the proposed method performs even better, with an accuracy of up to 87.19% and a recall of
up to 95.09%.